[None][test] add Kimi K3 feature matrix coverage

docs/source/models/supported-models.md

@@ -85,6 +85,7 @@ Note: Support for other models may vary. Features marked "N/A" are not applicabl
 | `Qwen4ExpForCausalLM`            | Yes               | Yes        | Yes                        | No                    | Yes             | Yes | No               | No                | No     | Yes           | Untested         | No             | N/A                      | Yes                   | Untested        |
 | `Llama4ForConditionalGeneration` | Yes               | Yes        | Yes                        | Yes                   | Yes             | No  | Yes              | Yes               | No     | Yes           | Yes              | Untested       | N/A                      | Yes                   | Yes             |
 | `GptOssForCausalLM`              | Yes               | Yes        | Yes                        | Yes                   | Yes             | No  | Yes              | No                | Yes    | Yes           | Yes              | Yes            | N/A                      | Yes                   | Yes             |
+| `KimiK3ForConditionalGeneration` [^15] [^17] | Yes               | Yes        | Yes                        | Yes                   | Yes             | No  | No               | No                | No     | Yes           | No               | Yes            | N/A                      | Yes                   | Yes             |
 | `Glm4MoeLiteForCausalLM` [^5]    | Yes               | Yes        | Untested                   | Untested              | Yes             | No  | No               | No                | No     | Yes           | Untested         | Untested       | N/A                      | Untested              | Untested        |
 | `NemotronHForCausalLM`           | Yes               | Yes        | Yes                        | Yes                   | Yes             | Yes | No               | No                | No     | Yes           | Yes              | Yes            | N/A                      | Untested              | Untested        |
 | `Gemma4ForConditionalGeneration` | Untested          | Yes        | Untested                   | No                    | Yes             | Yes | No               | No                | No     | Yes           | Untested         | No             | Yes                      | Untested              | Untested        |
@@ -107,6 +108,7 @@ Note: Support for other models may vary. Features marked "N/A" are not applicabl
 [^14]: Requires `transformers>=5.7.0`: MiniCPM-V 4.6 was upstreamed into transformers as a native model type (`minicpmv4_6`) and the checkpoint ships no remote code (`auto_map`) to fall back on. The Qwen3.5-hybrid text tower runs in BF16. Image, video, and text inputs are supported in this release (video reuses the same NaViT-packed vision path as image via `MiniCPMV4_6InputProcessor`).
 [^15]: Kimi K3 is only supported on NVIDIA Blackwell GPUs (`SM100` family); see the [Kimi K3 deployment guide](../deployment-guide/deployment-guide-for-kimi-k3-on-trtllm.md). Which recipes fit is set by the attention layout rather than by the GPU count: DEP16 (`enable_attention_dp: true`) replicates the BF16 non-expert weights on every rank (114 GB) on top of the MXFP4 routed experts at 16-way expert parallelism (90 GB), needing 210 GB per rank, so it requires GB300-class per-GPU memory. TEP16 (`enable_attention_dp: false`) shards those non-expert weights instead and needs 115 GB per rank; it is validated end-to-end on GB200 (`SM100`) at 16 GPUs. On B200 (`SM100`), kernel and module support is functional and covered by CI. Note that the FP8 weight-read path (TRTLLM-14765) does not relax the DEP16 requirement: the conversion runs after the weights are already resident in BF16, so it lowers the steady-state footprint but not the load-time peak.
 [^16]: Guided decoding for `Glm4MoeForCausalLM` is currently supported only with the `xgrammar` backend; `llguidance` is not working.
+[^17]: Kimi K3 supports suffix-automaton and DSpark speculative decoding. It has no MTP or EAGLE-3 head, and its DSpark checkpoints are not compatible with the plain `DFlash` mode represented by this matrix.
 
 # Encoder-Decoder Feature Support Matrix (PyTorch Backend)
 

tests/integration/defs/accuracy/accuracy_core.py

@@ -34,13 +34,39 @@
 from tensorrt_llm.logger import logger
 from tensorrt_llm.models.modeling_utils import QuantConfig
 from tensorrt_llm.quantization import QuantAlgo
+from tensorrt_llm.sampling_params import LogitsProcessor
 
 from ..common import venv_check_call, venv_mpi_check_call
 from ..conftest import llm_models_root
 from ..trt_test_alternative import check_call, exists
 from .video_mme import VideoMME as VideoMMEEvaluator
 
 
+class ForceTokenLogitsProcessor(LogitsProcessor):
+
+    def __init__(self, forced_token_id: int) -> None:
+        self._forced_token_id = forced_token_id
+
+    def __call__(
+        self,
+        req_id: int,
+        logits: torch.Tensor,
+        token_ids: list[list[int]],
+        stream_ptr: int | None,
+        client_id: int | None,
+    ) -> None:
+        del req_id, token_ids, client_id
+        if stream_ptr is None:
+            self._force_token(logits)
+            return
+        with torch.cuda.stream(torch.cuda.ExternalStream(stream_ptr)):
+            self._force_token(logits)
+
+    def _force_token(self, logits: torch.Tensor) -> None:
+        logits.fill_(float("-inf"))
+        logits[..., self._forced_token_id] = 0
+
+
 def compute_theta(num_samples: int,
                   sigma: float,
                   alpha: float = 0.05,
@@ -125,6 +151,23 @@ def assert_passing(self, accuracy: float) -> None:
             assert accuracy <= self.threshold, err_msg
 
 
+def compute_acceptance_length(llm: PyTorchLLM) -> float:
+    """Mean acceptance length over speculative iterations.
+
+    Requires enable_iter_perf_stats=True. Used by the AL-regression tests.
+    """
+    stats = llm.get_stats(timeout=2)
+    spec_iters = [
+        stat["specDecodingStats"] for stat in stats
+        if stat.get("specDecodingStats")
+        and stat["specDecodingStats"]["numDraftTokens"] > 0
+    ]
+    assert spec_iters, "No iterations with speculative decoding stats"
+    accepted = sum(stat["numAcceptedTokens"] for stat in spec_iters)
+    requests = sum(stat["numRequestsWithDraftTokens"] for stat in spec_iters)
+    return (accepted + requests) / requests
+
+
 def assert_acceptance_length(test_key: str, al_value: float) -> None:
     """Assert acceptance length meets the registered minimum.
 

tests/integration/defs/accuracy/test_glm52.py

@@ -17,7 +17,6 @@
 
 import jsonschema
 import pytest
-import torch
 
 from tensorrt_llm import LLM
 from tensorrt_llm.evaluate import JsonModeEval as JsonModeEvaluator
@@ -32,34 +31,9 @@
     SchedulingParams,
 )
 from tensorrt_llm.quantization import QuantAlgo
-from tensorrt_llm.sampling_params import LogitsProcessor
 
 from ..conftest import llm_models_root, parametrize_with_ids, skip_pre_blackwell, skip_ray
-from .accuracy_core import GSM8K, JsonModeEval, LlmapiAccuracyTestHarness
-
-
-class _ForceTokenLogitsProcessor(LogitsProcessor):
-    def __init__(self, forced_token_id: int) -> None:
-        self._forced_token_id = forced_token_id
-
-    def __call__(
-        self,
-        req_id: int,
-        logits: torch.Tensor,
-        token_ids: list[list[int]],
-        stream_ptr: int | None,
-        client_id: int | None,
-    ) -> None:
-        del req_id, token_ids, client_id
-        if stream_ptr is None:
-            self._force_token(logits)
-            return
-        with torch.cuda.stream(torch.cuda.ExternalStream(stream_ptr)):
-            self._force_token(logits)
-
-    def _force_token(self, logits: torch.Tensor) -> None:
-        logits.fill_(float("-inf"))
-        logits[..., self._forced_token_id] = 0
+from .accuracy_core import GSM8K, ForceTokenLogitsProcessor, JsonModeEval, LlmapiAccuracyTestHarness
 
 
 class _JsonModeGrammarEval(JsonModeEvaluator):
@@ -413,7 +387,7 @@ def _assert_logits_processor(llm: LLM) -> None:
                 max_tokens=output_length,
                 temperature=0,
                 end_id=-1,
-                logits_processor=_ForceTokenLogitsProcessor(forced_token_id),
+                logits_processor=ForceTokenLogitsProcessor(forced_token_id),
             ),
             use_tqdm=False,
         )